The full, user-facing history of Neurones 3D, grouped by engine domain. The v1.0 milestone is on the roadmap card at the top; v0.9 is the current launch candidate below it.
v1.0Coming soon
the milestone release, already taking shape in the workshop.
On the roadmap
Ultra-frugal video reconstruction: sharper, denser results with the comet-streak fuzz cleaned up, still fully on your own machine
A native macOS build, delivered as a signed .dmg next to the Windows executable
New one-click classifiers for outdoor, facade and infrastructure scenes
GPU-accelerated splat sorting for scenes beyond five million Gaussians
Faster still on very large clouds, with more of the index paged straight off disk
v0.36Coming soon: the drawing set release
the drawing set beside the scan, and video that knows when it has come back to where it started.
Plans, sections and elevations
Process > Floor plan... draws the real drawing on screen, through the same model the SVG, DXF and PDF writers use, at the sheet and scale it will be written at: what is on screen is the file
Sections and elevations: one command produces four elevations on the building's own axes and two sections across them, in all three formats
Cutting planes are placed by search, scored by the rooms and openings they cross, and a plane that would run along a wall is refused
The plan carries the section marks: the cut line, an arrow pointing the way you look, and the letter at each end
Exact hidden-line removal in every section and elevation, and an elevation shows the outside of the building
Colour only where the drawing already makes a distinction: outer walls darker than partitions, doors orange, glazing blue, each room washed in a faint colour of its own, and structure grey so the sheet still reads photocopied
A title block that carries its provenance: the maker's mark, the author, the date, a rule to sign on, and the sheet number and revision beside the size, the scale and the units
Hatched poche at 45 degrees, as ISO 128 draws cut material, honoured by the PDF, the SVG and the DXF alike and looking the same at 1:50 and at 1:500
The sheet obeys the standards it claims: lettering from the ISO 3098 series, the ISO 5455 scale list, and the ISO 5457 20 mm filing edge
Rooms are named only as far as the data supports: Corridor, Void, and on a classified cloud the dominant class in brackets, such as Room 3 (desk)
Better room detection: S3DIS Area 4 now resolves into 42 spaces where it was 16, and its largest is 279 m² rather than 513; a sprawling basin is called Open plan and is not washed
Cleaner drawings: dimension numbers sit above the line as read, external corners meet cleanly, walls meeting in a tee are joined, and doors and windows carry their own small dimension
Room labels stay clear of walls, of each other and of the dimensions, and never leave the building; every label, the north arrow's N included, sits inside the drawing frame
Room washes no longer overlap each other or leave a staircase along every edge
Room names in any script: the PDF embeds a font, so a Cyrillic title is drawn with zero question marks and stays searchable and selectable, and areas read m² instead of garbled characters
A dragged dimension or section mark that would land on something is refused, with a message
GIS and BIM
Vector layers, from the File menu or by drag and drop: GeoJSON, shapefile, DXF, SVG and CityGML land in the scan's own frame, drawn over the points in the viewport and in the 2-D plan tool, with their attributes on hover
A layer in a different coordinate system from the scan says so in orange
A plan sent out as SVG for markup comes back and sits over the scan it was drawn from, at the same scale
CityGML as features, a polygon per building with its height, year, function and the producer's own attributes, and LOD4 interiors as rooms that plan through the same writers
An IFC plans as a building rather than as a mesh, through the same SVG, DXF, PDF, GeoJSON and Shapefile writers as a survey, with nothing inferred
A scan knows its own coordinate system: read from LAS and LAZ, shown, and used as the default for a vector export
Exports keep the coordinate system they read, so a georeferenced survey no longer comes out unreferenced, and a cloud you have moved is not stamped with a reference it no longer has
3D reconstruction from video
Walk a room and come back, and the reconstruction now knows: places filmed twice are found and used to close the loop, where a minute of handheld footage had drifted 80° and folded the room onto itself
Views filmed far apart in time now agree 19x more often on that clip, from 0.7% to 13.2%
A loop is kept only if it measurably helps, so a wrong one is declined and clips with nothing to close come out exactly as before
One badly placed frame on a fast turn no longer bends the whole model: each frame votes, and a dissenting one is set aside
A black or blank video is an error, not a model: 30 s of black used to become 683,000 confident points, and now it says nothing was visible
Inspection and deliverables
A deviation is a set of defects, not just a percentage: failing points are grouped into regions with point count, centroid, bounding box, worst and mean deviation, worst first, and a click flies to each one
The punch list is also a column: colour, hide, section or export by region, where 1 is the first row of the list beside it
Conformance bands paint the deviation in four flat colours cut at one, two and four times the tolerance
When failures link into one region the size of the building, the report and the sidebar say so rather than list it as a finding
The punch list exports as CSV, including the zones that passed and the ones with too little data to judge, and the deviation exports as a column with the cloud
The report dialog can write the cloud beside the report, as LAS with its coordinate system on it
The registration is evidence: its RMS, inlier fraction and iteration count go into the report, and the reference system is stated, or marked "not stated in the source file"
A mesh says whether it is actually closed: boundary edges, non-manifold edges, shells, degenerate triangles and holes with their perimeter and area, shown in the mesh panel
Ground and change
The ground filter classifies the scene: Ground and Off-ground become classes of their own, with a checkbox to keep an existing classification instead
A groundify column, 1 for ground and 0 for everything else, painted in the class key's own colours as two blocks you can click off
Compare & detect change has its own place in the sidebar and in the Process menu
Carrying labels from a subsample back to the full cloud agrees with the source's own labels 97.64% of the time from a 1-in-10 decimation, 96.01% from 1-in-40 and 94.51% from 1-in-200
The terrain panel keeps its Build and write button on screen, however many options are open
The voxel view follows the colouring: change the colour mode and the cubes follow
Labelling and workspace
Labelling by class and shape: a shape row narrows within the chosen class, with counts inside that class
The labelling grid shows nine smaller tiles across, biggest components first under every filter, each in its component's own 3-D colour
Labelling keeps the component colours after each automatic save
Ctrl+click and right-click fly-to pick only what you can see, never a hidden component
Escape never closes or minimises the labelling tool
The labelled column is called n_classification, in the app and in exports
The Process menu has headed sections: Segment & label, Classify, Geometry, Plans & deliverables, Compare and Register
A compact sidebar: two lines of facts, fields as chips, one stacked bar per distribution
Isolating a class no longer thins the picture: the visible part gets the draw budget, up to 3x
Hiding one cloud no longer hides them all, and removing the opened cloud removes it from the viewport
Performance and reliability
A big E57 on Google Drive no longer sits at 0%: a Reading bar, a working cancel, and on a synced or network drive the first 1.5 M points on screen within seconds
Local E57 files stay fast: 1.6 s for a 127.8 M-point scan, from cache
The Process menu opens instantly
Toggling a class view on splats no longer freezes the interface
A half-downloaded LAS or LAZ says so instead of failing hundreds of megabytes later
A column computed for one scan no longer lingers in the Colour by list after loading another
v0.34Survey release
the outdoor half of the job, and models that open in a second.
Model engine
Open USD stages as first-class models: .usd, .usda, .usdc and .usdz, read by a native-Rust implementation with no pxr build and nothing extra in the installer
Composition resolved on load: references, payloads, sublayers, variant sets, class prototypes and native instancing
Verified against OpenUSD 26.8 on a 38-layer factory export: the same 65,654 meshes, the same 15,217,537 triangles, and a bounding box agreeing to the millimetre
metersPerUnit and upAxis honoured, including the common case where the root layer declares neither and every referenced layer does
A stage that takes 59 seconds to compose the first time reopens in 1.6 seconds, because the result is kept in a sidecar
Models arrive in named components: a tree in the scene panel, an eye on every component and every group
Double-click a model to fly to the component under the cursor, skipping the components you have hidden
Models are framed on the median and the 98th percentile, so one stray vertex cannot decide where the camera goes
A model dropped onto a scan is drawn as a model, not sampled into points
Export any model or cloud as a USD stage from the command line
Terrain and change
Fit the ground by cloth simulation, with no training data and no model file: F1 90.3% against a national mapping agency's own production labels, 16.2 million points in 2.14 seconds
Ground and height-above-ground land as scalar columns, exported to LAS as extra bytes and read back by PDAL, laspy and CloudCompare
The whole terrain deliverable in one pass off one ground fit on one grid: DTM, DSM, canopy height, slope, aspect, hillshade, contours, trees and clear-sky sunlight
Terrain rasters that leave the program: ESRI ASCII Grid and GeoTIFF, opened again in GDAL to check
Contours that cannot cross, and an interval finer than the data supports is warned about, with the finest defensible interval quoted
M3C2 change detection between two surveys, with a 95% level of detection and four verdicts: unchanged, gained, lost, and not observed
Carry a scalar column from one cloud onto another, 20 million points in 0.21 seconds
Intelligence
Four trained classifiers, and all of them ship in every build including the free one: Indoor over ten classes, Outdoor land cover, Powerlines and Streetscape
The classifier says how sure it was: the share of trees that agreed lands as a confidence column you can colour by
Primitive decomposition: 43.5 million points into 1,501 reviewable regions in 5.8 seconds
A semantic scene graph with the spaces that enclose the objects: 31 rooms and 48 openings on a test building in 0.35 seconds, exported as JSON, GraphML or CSV
Gaussian splatting
A level-of-detail tree for splats: 2.4x faster at a 500,000-splat budget, 4.6x at 150,000, on a 1.92 million splat capture
Rendering and workspace
Give every loaded cloud its own identity colour, so two surveys of one place stay apart
Demo scenes now come in pairs, so comparison and change detection can be tried with no data of your own
Performance and reliability
Slow formats are decoded once: USD, IFC and STEP results are kept in a sidecar, so every reopen is a memory-mapped read
A damaged USD file is refused rather than crashing the program
903 tests across 27 suites, plus a 17-case render QA matrix and a malformed-file corpus
v0.30Launch candidate
one Rust engine that takes you from capture to understanding to deliverable.
Point Cloud engine
Open a 43.5 million point LAS file in about 1.2 seconds with a memory-mapped, parallel reader, colour and intensity intact
Read compressed LAZ at roughly 40 million points per second, including COPC files
Load multi-scan E57 tripod captures with their poses applied and scans merged, plus the panoramic photos embedded inside
Open point-cloud PLY and NumPy NPY files, with any extra columns carried through as colourable fields
Work with files bigger than your RAM: an out-of-core index streams billions of points from disk, so the limit is your disk, not your memory
Reopen any indexed file in about a second, with almost no memory cost, because the index maps straight back in
Build the index for huge files on modest machines, at a bounded memory budget that stays roughly flat however large the file is
Drop several clouds into one scene as layers, each with its own visibility toggle
Colour by RGB, elevation, normals, intensity, classification or any scalar field, with a smooth ramp
Gaussian splatting
Render real 3D Gaussian Splatting natively, from PLY (INRIA) and SPZ (Niantic) files, auto-detected
A full EWA rasterizer with covariance projection, Mip-Splatting anti-aliasing and correctly sorted transparency
Navigate freely: splats re-sort in the background so movement never stutters, and sorting is now several times faster
Turn any ordinary point cloud into photoreal surfels: soft tangent-plane Gaussians that blend into continuous, seamless surfaces at full survey scale
Per-point normals are computed once and cached, so surfels and the normals view come back instantly on reopen
Rendering
Switch between crisp survey points and a continuous, lit, hole-filled surface with a single toggle
Eye-Dome Lighting gives raw points depth and readable silhouettes, the classic point-cloud look
Surface mode adds screen-space ambient occlusion and rim lighting for contact shadows and shape
Thin structures such as powerline wires stay crisp instead of melting into blobs
Oriented surfels tile dense scans into gap-free, matte surfaces without any per-frame sorting
Supersampling from 1x to 4x with contrast-adaptive sharpening for clean edges
Quality presets, Eco, Balanced and Premium, in one click, with integrated GPUs starting in Eco automatically
An occlusion-aware ground grid and a gradient backdrop composited in the same pass
A cinematic fly-in and fade reveal every time a scene loads
Mesh engine
Import textured glTF and GLB models, with base-colour textures, vertex colours and computed normals when missing
Open PLY meshes in binary and ASCII, with quads triangulated and smooth normals filled in when absent
Four shading modes, shaded, albedo, normals and flat, with a view-dependent headlight so form reads from any angle
Turn segmented point clouds into clean planar panels, ready to export
A built-in demo sculpture mesh so you can try the pipeline with no data on hand
BIM and CAD import
Drop an IFC building on the window and see it meshed by a pure-Rust importer, no external kernel and no conversion step
A 100,871 entity building with 3,956 elements meshes in about one second
Real solid modelling: extruded solids, faceted breps, mapped instances and true boolean clipping
Real materials: curtain-wall glass and mullions keep their IFC surface styles, and an architectural palette covers the rest
Files authored in millimetres land at the right size in the scene, every time
Import STEP and STP CAD parts, with exact B-rep tessellated at a size-relative tolerance for crisp curves
Buildings and parts composite in one scene alongside your point clouds and splats
Voxels
Turn any scan into a lit block world in real time, the Minecraft view, using the scan's own colours
Adjust resolution live, up to 2048 cells per axis
Export the voxel grid as glTF (GLB) for game engines and DCC tools
3D reconstruction from video
Drop a phone video on the window and get a colored 3D point cloud back in about fifteen seconds on a warm run
Feed-forward neural reconstruction runs locally on your own GPU, with no upload, no account and no cloud minutes
The model sizes itself to the VRAM you have, and falls back gracefully when acceleration or models are absent
Longer clips, up to around fifteen minutes, stay consistent through windowed gluing
Results come back upright automatically, levelled to gravity from the camera motion
Intelligence and classification
Label an indoor scan into floor, ceiling, wall and furniture in one click, at 91.9% accuracy on held-out data
A powerlines classifier sorts ground, vegetation, wire and pylon for utility corridors
A class legend lets you show, hide and isolate any category
Unsupervised segmentation decomposes any surface into coherent parts with no parameter to tune, plus a border-sharp fine preset
Per-segment analytics and per-segment meshing into clean planar panels
A benchmark-validated normals pipeline (bilateral PCA with gravity-aware orientation), cached for instant reuse
Toolbox: measurement, sections, floor plans, video
Measure point coordinates and 3D, horizontal and vertical distances, with octree-accurate picking under the cursor
Cut 2D floor plans and sections with a rotatable section plane and a live 3D slab preview, panned and zoomed like a CAD station
Export orthoimages up to 8192 px at full point density, as PNG or JPEG
Record a flythrough as you navigate, then replay it through a smoothing spline that keeps your pacing
Bake MP4 video of every render mode, from 720p to 1440p and 24 to 60 fps
Fly with WASD and a cruise-speed wheel, or orbit, pan and dolly
Save the current view as a 2x-resolution image
A gamified first-run tour with auto-completing steps, and three built-in demo scenes
Performance and reliability
60 fps on a 2,000,000,000 point cloud, out of core; a 43.5 million point scene measures about 110 fps at 1600x900
Pure Rust on GPU-native wgpu: no garbage collector, no runtime pauses
Out-of-core by design, so 10-year-old laptops stay interactive on files bigger than their RAM
A VRAM residency pool pages detail in and out, so clouds larger than your GPU never stall
Every optimization is measured against recorded baselines, including splat sorting several times faster and cached normals many times faster this release
A malformed-file torture corpus means corrupt, truncated or adversarial files can no longer crash the app
Crash logging and self-healing when the GPU runs out of memory
Distribution
One self-contained 15 MB Windows executable, no installer and no dependencies
A macOS build pipeline is ready
The free tier is the whole viewer and every in-app process, forever, with no trial clock and no account
Pro unlocks everything that writes results out: PLY and glTF exports, MP4 flythroughs and high-fidelity orthoimages
Pro is an offline, signed key validated on your own machine, with no server and no subscription, usable on any machine
One key covers every version released during its window
This site keeps only what it needs to work: a voting token on the board, a draft of a feedback message, and this choice. There is no analytics and no advertising. What is stored, and why.
